def get_intervention_results(templates,
tokenizer,
DEVICE='cuda',
gender='female',
layers_to_adj=[],
neurons_to_adj=[],
intervention_loc='all',
df_layer=None,
df_neuron=None):
if gender == 'female':
intervention_type = 'man_indirect'
else:
intervention_type = 'woman_indirect'
df = []
for template in templates:
# pickle.dump(template + "_" + gender, open("results/log.pickle", "wb" ) )
interventions = construct_interventions(template, tokenizer, DEVICE,
gender)
intervention_results = model.neuron_intervention_experiment(
interventions,
intervention_type,
layers_to_adj=layers_to_adj,
neurons_to_adj=neurons_to_adj,
intervention_loc=intervention_loc)
df_template = convert_results_to_pd(interventions,
intervention_results, df_layer,
df_neuron)
# calc odds ratio and odds-abs
df.append(df_template)
gc.collect()
return pd.concat(df)
def run_all(
model_type="gpt2",
device="cuda",
out_dir=".",
random_weights=False,
template_indices=None,
):
print("Model:", model_type, flush=True)
# Set up all the potential combinations.
professions = get_profession_list()
templates = get_template_list(template_indices)
intervention_types = get_intervention_types()
# Initialize Model and Tokenizer.
model = Model(device=device,
gpt2_version=model_type,
random_weights=random_weights)
tokenizer = (
GPT2Tokenizer if model.is_gpt2 else TransfoXLTokenizer
if model.is_txl else XLNetTokenizer if model.is_xlnet else
BertTokenizer if model.is_bert else DistilBertTokenizer if model.
is_distilbert else RobertaTokenizer).from_pretrained(model_type)
# Set up folder if it does not exist.
dt_string = datetime.now().strftime("%Y%m%d")
folder_name = dt_string + "_neuron_intervention"
base_path = os.path.join(out_dir, "results", folder_name)
if random_weights:
base_path = os.path.join(base_path, "random")
if not os.path.exists(base_path):
os.makedirs(base_path)
# Iterate over all possible templates.
for temp in templates:
print("Running template '{}' now...".format(temp), flush=True)
# Fill in all professions into current template
interventions = construct_interventions(temp, professions, tokenizer,
device)
# Consider all the intervention types
for itype in intervention_types:
print("\t Running with intervention: {}".format(itype), flush=True)
# Run actual exp.
intervention_results = model.neuron_intervention_experiment(
interventions, itype, alpha=1.0)
df = convert_results_to_pd(interventions, intervention_results)
# Generate file name.
temp_string = "_".join(temp.replace("{}", "X").split())
model_type_string = model_type
fname = "_".join([temp_string, itype, model_type_string])
# Finally, save each exp separately.
df.to_csv(os.path.join(base_path, fname + ".csv"))
def main():
DEVICE = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
tokenizer = BertTokenizer.from_pretrained('bert-base-cased') # changed
model = Model(device=DEVICE)
base_sentence = "The {} said that"
biased_word = "teacher"
intervention = Intervention(tokenizer,
base_sentence, [biased_word, "man", "woman"],
["he", "she"],
device=DEVICE)
interventions = {biased_word: intervention}
intervention_results = model.neuron_intervention_experiment(
interventions, 'man_minus_woman')
df = convert_results_to_pd(interventions, intervention_results)
print('more probable candidate per layer, across all neurons in the layer')
print(df[0:5])
def run_all(
model_type="gpt2",
device="cuda",
out_dir=".",
grammar_file=None,
structure=None,
random_weights=False,
template_indices=None,
):
print("Model:", model_type, flush=True)
# Set up all the potential combinations.
if grammar_file is not None:
if structure is None:
raise Exception(
"Error: grammar file given but no structure specified")
grammar = read_grammar(grammar_file)
professions = get_profession_list(grammar=grammar, structure=structure)
templates = get_template_list(structure=structure, grammar=grammar)
intervention_types = get_intervention_types(bias_type="structural")
else:
professions = get_profession_list()
templates = get_template_list(template_indices)
intervention_types = get_intervention_types()
# Initialize Model and Tokenizer.
tokenizer = GPT2Tokenizer.from_pretrained(model_type)
model = Model(device=device,
gpt2_version=model_type,
random_weights=random_weights)
# Set up folder if it does not exist.
dt_string = datetime.now().strftime("%Y%m%d")
folder_name = dt_string + "_neuron_intervention"
base_path = os.path.join(out_dir, "results", folder_name)
if random_weights:
base_path = os.path.join(base_path, "random")
if not os.path.exists(base_path):
os.makedirs(base_path)
# Iterate over all possible templates.
for number in ('sing', 'pl'):
for temp in templates[number]:
print("Running template '{}' now...".format(temp), flush=True)
# Fill in all professions into current template
if structure == "simple_agreement":
other_number = 'sing' if number == 'pl' else 'pl'
interventions = construct_interventions(
temp,
professions[number],
tokenizer,
device,
structure=structure,
number=number,
subs=professions[other_number])
else:
interventions = construct_interventions(temp,
professions[number],
tokenizer,
device,
structure=structure,
number=number)
# Consider all the intervention types
for itype in intervention_types:
print("\t Running with intervention: {}".format(itype),
flush=True)
# Run actual exp.
intervention_results = model.neuron_intervention_experiment(
interventions, itype, alpha=1.0)
df = convert_results_to_pd(interventions, intervention_results)
# Generate file name.
temp_string = "_".join(temp.replace("{}", "X").split())
model_type_string = model_type
fname = "_".join([temp_string, itype, model_type_string])
# Finally, save each exp separately.
df.to_csv(os.path.join(base_path, fname + ".csv"))